fix(study): the typing card — keyboard on the front, mid-flip stutter, and a card that jumped

[jvsena42]

Four fixes to the study screen on a deck with Type the answer on. Each was measured on emulator-5554; the numbers are in journeys/RESULTS.md.

1. The keyboard came up over the next card's front

AnimatedContent keeps the outgoing card composed for its fade and re-runs the content lambda while it does, so the input read state.typePhase — by then already Answering for the card that had just arrived. The outgoing card drew the incoming card's input and its FocusRequester fired. CardSnapshot now carries the card's own phase and typed text.

8 of 8 advances → 0 of 19.

2. The turn stuttered halfway through

Not the card. The field took focus the instant the back face composed — the frame the rotation crosses 90° — and showing the IME makes SurfaceFlinger allocate its window surface. atrace over one flip:

460 ms  putmethod.latin   Choreographer#doFrame
446 ms  putmethod.latin   draw-VRI[InputMethod]
426 ms  RenderThread      dequeueBuffer / allocateHelper
433 ms  surfaceflinger    present
425 ms  jvsena42.loopky   eglSwapBuffersWithDamageKHR   <- us, blocked

The field now waits out the turn before asking for focus. Over the 750 ms of a flip: 21 frames → 44 (it was dropping every other one), janky 13.6% → 7.1%, p90 32 ms → 16 ms. The IME's ~445 ms is unchanged and unavoidable — it is now spent where it reads as the keyboard arriving.

Seven other explanations were built and benched first, and all seven were wrong: a height animation on the rows below the card, composing the back face at the crossing, TextAutoSize, the rounded clip outside the 3D layer, rotationY itself, the reveal haptic, the card's ripple.

3. The card jumped when the grades appeared

The card is weight(1f) of a column the screen padded for the IME, so the keyboard's departure resized it — right as a checked answer brought the grades in. Its y ran 405 → 579 → 405 across front / answering / graded, every word re-flowing through TextAutoSize on the way. That is the "flash": no dropped frames (3.2% janky either way), the whole layout moving in one frame.

The screen no longer pads for the IME. The card keeps its height and its lower edge passes behind the keyboard; the input block sits at the card's middle, clear of it (controls at 953..1434 against a keyboard top of 1524), and still scrolls where it would not be.

Card y after: 405..1875 in all three states. Same in landscape at w914dp (332..869), which also makes typing usable there for the first time — the card had been a ~40 px strip with its controls unreachable.

4. Two hard cuts, softened

The miss line grows in rather than appearing under the finger already reaching for Check.

Verified

Journey 03, extended with a typing block and the two checks these fixes would have failed (frames per turn; study_card bounds identical across states). Portrait and landscape, detektAll and assembleDebug green. journeys/RESULTS.md records the runs, including the control experiment — tab switches at 1–4% janky against the flip's 13% — that turned this from guessing into measuring.

🤖 Generated with Claude Code